New framework transfers LLM safety from high-resource to low-resource languages

Researchers have developed a new framework called Multilingual Self-Distillation (MSD) to improve the safety alignment of large language models (LLMs) in low-resource languages. This method transfers safety capabilities from high-resource languages, like English, to others, such as Javanese, without requiring specific safety data for each target language. The framework utilizes multilingual queries and a novel optimization technique called Dual-Perspective Safety Weighting (DPSW) to enhance cross-lingual safety transfer while maintaining general model capabilities. AI

IMPACT This research could lead to more robust and equitable AI safety across diverse languages, reducing vulnerabilities in low-resource settings.

RANK_REASON This is a research paper detailing a new framework for improving LLM safety alignment. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• Dual-Perspective Safety Weighting
• LLMs
• Multilingual Self-Distillation

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →